💰 quantitative finance

Modeling Story Expectations: A Generative Framework using LLMs

This paper introduces a generative framework using large language models to extract interpretable features from imagined story continuations, demonstrating that these LLM-derived expectations effectively predict both human beliefs and actual narrative outcomes while significantly correlating with reader engagement.

Original authors: Hortense Fong, George Gui, Bo Yang

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hortense Fong, George Gui, Bo Yang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're sitting on the couch, glued to a TV show or a book, totally hooked. You're not just watching what's happening right now; you're constantly guessing what's going to happen next. Will the hero survive? Will the villain get caught? Will the romance bloom? That feeling of "what's next?" is a huge part of why we keep reading or watching. But for a long time, scientists and storytellers have had a hard time measuring those guesses. You can't easily ask a million people to write down every single possibility they imagine for the next chapter. It's like trying to catch smoke with your hands.

Enter this paper, which acts like a digital crystal ball. The authors, Hortense Fong, George Gui, and Bo Yang, decided to use a super-smart computer brain (called a Large Language Model, or LLM) to do the guessing for us. Think of the LLM as a super-avid reader who has swallowed the text of over 125,000 books. Because it has read so much, it knows how stories usually go.

The Magic Trick: Dreaming Up the Future
Here's how they did it. They took a story up to a certain point (say, the end of Chapter 1) and asked the computer: "Okay, imagine 50 different ways this story could continue in Chapter 2." The computer didn't just write one ending; it wrote 50 slightly different versions, creating a whole cloud of possibilities.

From these 50 imagined futures, the researchers didn't just look at the words; they pulled out specific "flavors" or features. They looked at things like:

Valence: Is the next chapter going to be happy or sad?
Arousal: Is it going to be high-energy and exciting, or low-energy and calm?
Speed: Will the plot move fast or slow?
Circuitousness: Will the story take a straight path or a winding, twisty road?

They treated these 50 imagined chapters like a crowd of people voting on what they think will happen, and then they averaged the votes to get a "prediction" of what a typical reader expects.

Did the Computer Get It Right?
The big question was: Does this computer guessing game actually match what real humans think? The authors tested this in two ways.

First, they ran a controlled experiment with 263 real people. These humans read short stories and were asked, "What do you think the next chapter will feel like?" The researchers compared the humans' answers to the computer's "50 imagined chapters."

The Result: The computer was surprisingly good. When the humans said they expected a high-energy chapter, the computer's imagined chapters were also high-energy. The connection was strong, though not perfect (the computer's predictions were a bit more "compressed" or less extreme than the humans' wild guesses).
The Catch: The computer was great at guessing the average feeling, but it wasn't as good at guessing how uncertain people were. Humans have a wide range of doubts; the computer's guesses were a bit too confident and concentrated.

Second, they looked at real data from an online reading platform with 937 books and 8,399 chapters. Here, they couldn't ask readers what they thought, so they used a logic trick called "rational expectations." This assumes that if you know how stories work, your guesses should match what actually happens in the end. They compared the computer's imagined chapters to the actual chapters that were eventually written.

The Result: The computer's predictions were informative predictors of what actually happened. If the computer imagined a fast-paced next chapter, the real next chapter was often fast-paced. This held true for emotions, speed, and even complex things like "tension" or "cliffhanger intensity."

Why It Matters: The "Next Chapter" Factor
The most exciting part of the paper is what they found when they looked at how much people engaged with the stories (like clicking "continue," voting, or commenting).

They discovered that knowing what a reader expects to happen next is just as important as knowing what they just read.

In the human survey, if a reader expected the next chapter to be arousing (high energy), they were much more likely to say they wanted to keep reading. This was true even if the chapter they just finished was calm.
In the real-world data, the same thing happened. If the computer predicted a high-energy or fast-paced next chapter, readers were more likely to vote and comment.

The paper explicitly rules out the idea that we only care about what's happening right now. It suggests that our "forward-looking beliefs" (our guesses about the future) drive our behavior. It's not just about the current scene; it's about the promise of the next one.

What the Paper Does NOT Say
It's important to know what this isn't. The authors are very careful not to say they have solved the mystery of human behavior or that the computer is a perfect mind-reader.

They do not claim the computer's predictions are identical to human thoughts. The computer's guesses are "imperfect but informative." They capture the direction and the general vibe, but they miss some of the nuance and uncertainty that real humans feel.
They do not say this proves that changing a story's ending causes more engagement. They found a strong link (association), but because stories are complex and change many things at once, they can't say for sure that one specific feature is the sole cause. They suggest this method helps researchers ask better questions for future experiments.
They do not claim this works for every type of media yet. Their study focused on text (books and stories). They admit that movies and music have visual and sound elements that their text-based computer brain can't see or hear yet.

The Bottom Line
This paper suggests that we can use a super-reader computer to model what we, as an audience, are thinking about the future of a story. It's a new, scalable way to peek into the "what if" zone of our brains. By understanding what readers expect to happen next, creators and platforms might finally understand why some stories are page-turners and others are put down halfway through. It's not a magic wand that solves everything, but it's a powerful new tool for understanding the story we tell ourselves before we even turn the page.

Technical Summary: Modeling Story Expectations with Large Language Models

Problem Statement

Consumers' engagement with narrative media (books, films, TV) is significantly shaped by their forward-looking expectations regarding what will happen next. However, modeling these expectations over unstructured narrative content has historically been challenging. Traditional approaches face two main hurdles:

Survey Limitations: Eliciting a full distribution of possible story continuations from human respondents is cost-prohibitive, while surveying only one or two dimensions fails to capture the complexity of narrative unfolding.
Modeling Assumptions: Standard economic models often assume rational expectations over low-dimensional outcomes (e.g., future prices). Applying this to stories requires modeling how a high-dimensional distribution of possible continuations depends on a high-dimensional history of the story, for which there is no obvious starting point.

Consequently, there is a lack of scalable methods to approximate consumer beliefs about narrative content, limiting the ability of creators and platforms to understand how anticipated content drives engagement.

Methodology

The authors propose a generative framework utilizing Large Language Models (LLMs) to approximate consumers' forward-looking beliefs. The framework consists of four components:

1. Imagined Story Continuation Generation

The core mechanism involves using a pre-trained LLM to generate multiple plausible continuations of a story given the text read so far (e.g., chapters 1 through $t$ ).

Process: The LLM is prompted to generate the next chapter ( $t+1$ ). To approximate a distribution of beliefs rather than a single deterministic output, the model generates $N$ continuations (typically 50) per chapter.
Parameters: The authors use a temperature of 1.0 to balance diversity and coherence.
Rationale: LLMs, trained on vast corpora of narrative text, capture complex storytelling patterns and can generate a distribution of continuations that mirrors the expectations of a representative consumer.

2. Feature Extraction

The unstructured text of the imagined continuations is converted into interpretable, theory-motivated features.

Features: The framework extracts predefined features associated with engagement, such as emotion (valence, arousal) and semantic path (speed, circuitousness). It also extracts "learned" narrative dimensions (e.g., tension, curiosity, cliffhanger intensity, pacing).
Expectation Calculation: The expectation for a feature $z$ at chapter $t$ is calculated as the average of the feature values extracted from the $N$ imagined continuations:
$E_t^n[e_{i,t+1}^n] = \frac{1}{N} \sum_{n=1}^N e_{i,t+1}^n$
Higher-order moments (variance, deviation) can also be calculated, though the paper primarily focuses on the first moment (expectation).

3. Validation

The authors propose two complementary validation procedures to assess whether LLM-derived expectations are valid proxies for human beliefs:

Survey-Based Validation: Compares LLM-derived expectations against human-reported beliefs collected via surveys. Respondents report probability distributions over feature outcomes (e.g., valence bins) for the next chapter.
Rational-Expectations Validation: Compares LLM-derived expectations against the actual realized story continuations. Under the assumption that consumers (and the LLM) understand how stories typically unfold, the LLM's predicted distribution should align with the objective distribution of actual story outcomes. This is tested by regressing actual feature values on LLM-derived expected values.

4. Engagement Analysis

The framework is used to model reader engagement (interest, decision to continue, voting, commenting) as a function of both realized content (features of the chapter just read) and expected content (features of the imagined next chapter). The authors employ a linear expected utility model where engagement depends on current features and the expectation of future features.

Empirical Applications and Results

The framework was tested in two settings: a controlled survey experiment and an observational study of an online reading platform.

1. Survey Data (Controlled Setting)

Data: 263 participants read short stories across five genres and reported emotional beliefs (valence/arousal) and engagement intentions.
Validation: LLM-derived expectations for valence and arousal significantly correlated with human-reported expectations (slopes of 0.57 and 0.56, respectively), though LLMs tended to compress the range of beliefs.
Engagement Findings:
- Stated Interest: Expected arousal significantly predicted interest in continuing to read, even after controlling for current chapter valence and arousal. Expected valence was not significant.
- Decision to Continue: Expected arousal was a significant predictor of the binary decision to read the next chapter.
- Proxy Performance: Replacing survey-elicited expectations with LLM-derived expectations in regression models yielded comparable results, confirming LLMs as a scalable proxy.

2. Observational Data (Online Platform)

Data: 8,399 chapters from 937 books on an online platform, with engagement measured by vote-to-read and comment-to-read rates.
Validation (Rational Expectations): LLM-derived expectations for 15 features (including emotion, semantic path, and narrative dimensions) were significant positive predictors of actual chapter features. The coefficients were positive and significant, though attenuated (e.g., valence slope 0.69, arousal slope 0.62).
Engagement Findings:
- Forward-looking expectations (valence, arousal, speed, circuitousness) were significant predictors of voting and commenting rates, above and beyond the features of the content already consumed.
- Specifically, more negative/low-energy expectations and higher speed/lower circuitousness expectations were associated with higher engagement.
- Sensitivity Analysis: The authors demonstrated that the coefficients for focal features (e.g., expected arousal) are sensitive to the inclusion of other narrative dimensions (e.g., tension, cliffhangers), highlighting the importance of the framework in refining causal questions.

Key Contributions

Scalable Framework for Unstructured Expectations: The paper introduces a method to model consumer beliefs over high-dimensional, unstructured narrative content, extending the expectations literature beyond structured state spaces.
Dual Validation Strategy: It proposes and implements two distinct validation approaches (survey-based and rational-expectations-based), demonstrating that LLMs can serve as informative proxies for human beliefs without requiring costly data collection in all contexts.
Forward-Looking Engagement Drivers: The study provides empirical evidence that what readers expect to happen (specifically expected arousal and semantic path features) drives engagement independently of what they have already read.
Feature Discovery and Causal Refinement: By generating continuations, the framework allows researchers to extract a wide range of narrative dimensions and test how controlling for these dimensions affects the estimated impact of focal features, thereby helping to diagnose omitted variable bias in narrative studies.

Significance and Claims

The authors claim that this framework provides a scalable method for modeling consumer beliefs about narrative content. They assert that:

LLMs trained on narrative text can serve as a "reasonable starting point" for modeling reader anticipation.
Forward-looking expectations are a distinct and significant driver of engagement, suggesting that content creators and platforms should consider how narrative choices shape future expectations, not just immediate reactions.
The method complements existing content analysis by adding a forward-looking dimension, which is crucial for understanding dynamic behavior in narrative media.

The paper remains modest regarding causality, noting that their results are associational and that the LLM-derived expectations are approximations (introducing measurement error). They acknowledge limitations regarding the heterogeneity of readers (treating the LLM as an "average" reader) and the text-only nature of their current implementation, suggesting future work could explore multimodal inputs and heterogeneous consumer segments.